Grounded Generation with Citations

You are a research assistant. Answer the question using only the provided documents. Every factual claim in your answer MUST include an inline citation in the format [Doc N].

Documents:
[Doc 1]: {{DOCUMENT_1}}
[Doc 2]: {{DOCUMENT_2}}
[Doc 3]: {{DOCUMENT_3}}

Rules:
1. Cite the document number inline for every claim: e.g., "The model was released in 2023 [Doc 2]."
2. Do not add information not present in the documents.
3. If the documents do not contain enough information to answer the question, respond with:
   "The provided context does not contain enough information. Consider retrieving documents on: [2-3 specific topics that would help answer the question]."
4. End your response with a Sources section listing each document you cited.

Question: {{USER_QUERY}}

Sources:

What this prompt does

Requires the model to attach a document citation to every factual claim, surfaces a structured gap analysis when context is insufficient, and lists all cited sources at the end. Builds verifiable, auditable RAG responses.

When to use it

• RAG applications where users need to verify claims against source documents
• Enterprise search where answer provenance is a compliance requirement
• Any pipeline where you must distinguish retrieved facts from model inference

When NOT to use it

• Conversational chat without document retrieval — citation format adds clutter
• Creative or generative tasks where attribution is not meaningful

The Prompt

Why it's structured this way

Numbered documents [Doc N] — Named labels (e.g., "Finance Report Q3") work but numbered labels are shorter and less likely to be paraphrased incorrectly by the model.

"MUST include an inline citation" — Without explicit enforcement, models cite selectively. "MUST" in caps increases citation recall on every claim.

Gap analysis with topic suggestions — A bare "I don't know" is not actionable. Asking the model to name 2–3 missing topics tells the retrieval system exactly what to fetch next.

Sources section at the end — Provides a machine-parseable reference list separate from inline citations, useful for building UI link panels.

Variations